Evaluating Ensemble Kalman, Particle, and Ensemble Particle Filters through Soil Temperature Prediction

Abstract

Data assimilation is a useful tool in hydrologic and agricultural application studies because of its ability to produce predicted results with high accuracy. However, different data-assimilation methods have different performances for a given application. Although the popular ensemble Kalman filter (EnKF) performs well with Gaussian distribution, the error is difficult to conform to the Gaussian distribution. To take advantage of the EnKF, this study presents a new data-assimilation method, ensemble particle filter (EnPF), which is an integration of the EnKF and the particle filter (PF). This new method was evaluated in comparison with two existing methods (EnKF and PF) through soil temperature predictions. The simple biosphere model (SiB2) and the filters were assessed with observations from the Wudaogou experimental area in the Huaihe River basin, China. Results show that when the time interval increases adequately, all the simulated or assimilated results improve significantly. All of these filters tend to be more stable when the number of particles reaches a certain amount (e.g., 60) or the variance is small (e.g., less than 0.6) in the study. When the number of particles is less than a threshold value (e.g., 30), the advantage among these three methods is not appreciable. The error obtained by EnPF is smaller than that by EnKF and PF; this means that EnPF performs better than EnKF and PF.